POLICY RESEARCH WORKING PAPER 3 010 Vulnerability in Consumption, Education, and Health Evidence from Moldova during the Russian Crisis Edmundo Murrugarra Jose Signoret The World Bank Europe and Central Asia Region Human Development Sector Unit April 2003 I POLICY RESEARCH WORKING PAPER 3010 Abstract Murrugarra and Signoret analyze the widespread effects release labor supply. Health utilization decreased mainly of the financial crisis in Russia to explore the for primary health care (not for hospitals), both for vulnerabilities of households in Moldova. They better-off households and in rural areas. Some of these show that the crisis had differential impacts on changes are due to limited household resources (health), households, affecting most the urban and better-off. decreased public spending (health and education) or the Households' decisions about education and health need to increase households' labor supply (education of resulted in decreased utilization and expenditures. The teenagers). Social benefits played a very limited role in enrollment of young children from better-off households mitigating these effects, solely in health care use. did not improve while others did. Secondary school Households' assets helped to offset some of the negative enrollment of children from better-off households effects of declining incomes. decreased after the crisis, in part because of the need to This paper-a product of the Human Development Sector Unit, Europe and Central Asia Region-is part of a larger effort in the region to understand the effects of crises on human development. Copies of the paper are available free from the World Bank, 1818 H Street NW, Washington, DC 20433. Please contact Edmundo Murrugarra, mail stop MC7-703, telephone 202-473-4452, fax 202-477-3387, email address emurrugarra@aworldbank.org. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. Jose Signoret may be contacted at signoret@econ.berkeley.edu. April 2003. (46 pages) The Polcy Researcb Working Paper Seues d bssem yates the ftdegs of work A i progress to encourage the exchange of adeas about development issues. An objective of the series is to get the findings oiit qiickly, evenz If the presentations are less than fully polisbed. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed fi this paper are enttrely those of the auithors They do not necessarily represent the view of the World Bank, its Executive Directors, or the countries they represent. Produced by the Research Advisory Staff Vulnerability in Consumption, Education and Health: Evidence from Moldova during the Russian Crisis' Edmundo Murrugarra a Jose Signoret b ' Paper prepared for the Non Income Dimensions of Poverty Regional Study. a The World Bank. Human Development Department, Europe and Central Asia Region. e-mail: emurrugarra(worldbank.org. b U. of California, Berkeley, Dept. of Economics. e-mail: signoretXecon.berkeley.edu Vulnerability in Consumption, Education and Health: Evidence from Moldova during the Russian Crisis 1. Introduction This paper provides an assessment of household vulnerability in consumption in Moldova and discusses three major questions: (i) Which households were most affected by the effects of the Russian crisis in Moldova; (ii) How households characteristics helped households in coping with the shock; and (iii) Did government and informal safety nets provide a support during times of crisis? Vulnerability is defined here as welfare losses measured as consumption drops. Households, however, may be vulnerable to a number of shocks from different nature. Unexpected events in labor markets (unemployment), productive activities (input prices or access to markets), weather and natural phenomena (earthquakes, droughts, floods), or aggregate macroeconomic shocks (exchange rate devaluations). This paper exploits the effects of the financial crisis in Russia to examine the impact on poverty and consumption levels in Moldova. Moldova is strongly linked to the Russian economy through commercial, demographic (migration) and historical reasons. These links made the Moldovan economy and society particularly sensitive to the events in the Russian economy. The Russian financial crisis, though short-lived, significantly impacted through the devaluation of the ruble and the reduced exports to Russia. The rest of the paper is organized as follows. Section 2 describes the timing and nature of the Russian financial crisis and the effects on the Moldovan economy. Section 3 presents the measures of vulnerability and discusses methodological approaches used in the literature. Section 4 provides the results from the estimation and discusses the results. Section 5 summarizes the results. 2. Macroeconomic context and the Russian crisis Moldova is a landlocked poor country of less than 4 million population, with a national GDP per capita around US$ 400 but subject to significant economic shocks during the last decade. Despite the significant declines of GDP after independence in the early nineties, GDP per capita started recovering until 1997, only to suffer the effects from the Russian crisis that reduced GDP by more than 6 percent in 1998. By 1999, GDP per capita was similar to that of 1994 (Moldova Economic Trends, 2001). Similar to other former Soviet Union republics, Moldova was also characterized by its barter economy, in particular with the central partner (Russia). After the collapse of the Soviet Union, Moldova (as other FSU countnes) continued having strong commercial attachments with Russia and did not diversify their commercial partners. Compared to other economies, the ratio of exports and imports to GDP is relatively large in Moldova. During the nineties, exports represented more than 50 percent of GDP in Moldova, and imports more than 70 percent. In 1997, the trade intensity ratio (exports plus imports divided by GDP) was 1.29, one of the largest among FSU countries with the exception of Tajikistan (1.6) and close to Belarus (1.26).' Table 1: Source: Moldovz Economic Trends (2001) Tabls 1.1 GDP and GDP par Capft; 1993 - 28X 199 I2 196 In 1 1" o It" 203 prl Nominal ", Lei mUn1n 1,821.1 4,736.8 6,479.7 7,7976 8,917.0 9,1221 12,321.6 15,9800 Real GDP, year-on-year % chaMe -1.2 -30.9 -1.4 .59 1.6 6.5 -3.4 1.9 Populson, thousds 3,607.6 3,60.5 3,603.7 3,5090 3,664.0 3,648.3 3,645.3 3,63s.5 A gera EsMane Rate (Lei I USX 1.45 4.0S 4.49 4.59 4.63 5.38 10.51 12.43 E,tn8 Rate, end of pariod (Ld I USS) 3.64 4.27 4.50 4.65 4.66 8.32 11.59 12.38 GDP per cap'a (US$.acurlt pims) 348.1 323.3 400.5 472.0 527.1 48.8 321.6 353.5 Source: Depatmed Stof and Scdoiogy, MET calculat*n Furthermore, the trade intensity of the Moldova economy was concentrated on a key partner, Russia, and on other FSU countries. In 1997, about 70 percent of the exports and 52 percent of the imports were with CIS countries. In addition, besides being the most important commercial partner, Russia is also the most important direct investor (Economic Trends, 2001). TabBe: Trade linntensity an CI[S couintries 1996-2001 Country Name 1996 1997 1998 1999 2000 2001 Armenia 79.2 78.5 71.8 70.6 74.1 72.3 Azerbaijan 85.1 82.0 77.1 69.9 79.1 78.5 Belarus 96.8 125.5 123.0 120.8 137.2 82.0 Bulgaria 122.7 118.3 99.0 96.0 122.5 125.1 Croatia 89.9 97.9 88.8 89.4 95.6 101.6 Georgia 46.0 56.7 53.4 57.2 63.4 60.3 Hungary 78.8 91.0 103.3 108.5 126.5 123.1 Kazakhstan 71.3 72.4 65.2 82.6 106.0 98.4 Kyrgyz Republic 87.3 84.5 94.5 99.2 89.4 74.2 Moldlova 129.2 129.1 119.0 118.7 126.5 124.9 Romania 64.7 65.4 54.6 61.7 73.6 72.5 Russian Fed. 45.5 45.5 57.3 71.1 68.6 61.0 Tajikistan - 168.5 103.1 126.4 165.4 - Turkmenistan 150.0 101.6 94.2 103.5 116.4 - Ukraine 93.9 84.2 86.0 102.0 120.4 110.1 Uzbekistan 61.9 57.0 45.3 36.6 46.1 84.8 Source: SIMA database. As a comparison, a study using data for 1982 (Leamer, 1988) showed that only Singapore (1.62), French Guyana (1.25) and Brunei (1.07) have high intensities. In the study, Hungary was the only East European or non-market economy with an intensity of 0.06. 3 Trade intensity ratios may not measure precisely the degree of openness of an economy since other factors -- such as resources, prices, tastes and even natural barriers to trade - may also affect the level of exports and imports compared to GDP (Leaner, 1988). With this caveat in mind, trade intensity was associated an increased vulnerability of the country to shock in Russia as it is suggested in Figure 1. Countries with higher trade intensity ratios experienced a more rapid decrease in growth rates between 1997 and 1999. Growth rates decreased between 5 to 8 percentage points in Moldova and Belarus, countries with large openness indexes. Less open countries like Uzbekistan and Romania did not experience this decrease in growth rates. Notice that this evidence does not attempt to discuss the relationship between openness and growth, because we examine trade intensity (not exactly openess) and reduced growth rates in a context of concentrated trade and a financial crisis. Openess and slower growth during the Russian Crisis 7 0- 5.0 RU 'RO ~3 0 UK *UZ AZ 1 0 - KA *'AR * HU -1.0 o-3.0 -5.0 *MD -7 0 - G BHR GE BL -9.0 40 60 80 100 120 140 Openess Index (M+X / GDP) Figure 1 The financial crisis in Russia forced the devaluation of the Russian Ruble against the US Dollar. While the crisis was short lived and Russia soon experienced recovery, the permanent changes in relative prices (exchange rates) affected in a more persistent fashion the former FSU economies. For the FSU economies, the Ruble devaluation represented a persistent appreciation of local currencies against the Ruble. 2 In fact, it would be interesting to exanune the evidence from FSU countnes to analyze the lmkage between openness and growth during cnses. 4 Moldova Lei-R.Ruble exchange rate l _ ___ _ - 160 09 _ISO__ 150 0 8 _ \ \__-140 0713 0.6 120_ ) 0 5 _ __1 =0_ l 0 4 100 0.3 l l l l l l l l l90 Source Moldova Econonnc Trends (2001) IFigure 2 Event though the nominal and real Lei/Ruble rates were slowly falling since 1996, during the third and fourth quarter of 1998 both fell more significantly. The nominal rate (continuous line) fell 0.78 Lei per Ruble in the preceding months to 0.40 during the fourth quarter of 1998 and even to 0.37 in the first quarter of 1999 (almost a 50 percent reduction). It would later reach an equilibrium level at 0.44 afterwards representing a 44 percent decrease. The real rate (dashed line) followed the same pattern, although less pronounced because of the lagged inflationary effect on non-food items and services. It decreased more than 21 percent when comparing the fourth quarter with the second quarter of 1998. In the short run Moldova experience also the negative income effect in Russia (recession) that outweighed the competitiveness gains. A more important implication was the surge in the Lei/Dollar exchange rate. Even though the Lei lost value against the US dollar during most of the decade, this depreciation was very gradual: during 1997 it only increased 0.7 percent. Between August and November 1998, however, the Lei/USD exchange rate increased by 60 percent, and continued increasing during the fourth quarter (almost 80 percent compared to August 1998). The inflationary consequences of the Lei/Dollar increase were reflected in the subsequent months as seen in Figure 3. In summary, even though the economy of Moldova was still suffering the effects of the recession during the late nineties, the Russian financial crisis negatively affected the economy through price increases due to the Lei/Dollar devaluation, temporary loss of Russian external markets due to recession, and other intermediate effects (such as decreased remittances from Russia). 5 Moldova: Prices and Lei/USD exchange rate Soo - _ __ _ _ __ _ _ 13 450 12 400 12 35s) 350 / ,t' '10 300 250 - 200 - - - O - - - - - - - - - - - - - - - - - - - - - 7 lo: ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~6 so 5 0 . . .4 - - Consumet Pnce Index (CPI), Dec-94-100 Total CPI - - - Wholesale Price Index (WPI), Dec-94-100 Exchange rate, LeL/USS Source Moldova Economic Trends (2001) Figure 3 This paper exploits the observed effects of the Russian crisis on the Moldovan economy to identify the vulnerability in consumption, health and education among households. The next section discusses those dimensions of development and their methodological approach. 3. Measures of welfare and vulnerability This paper uses the definition of vulnerability as "the propensity to suffer a significant fall in welfare." The operational representation involves two different dimensions. First, the measure of welfare requires narrowing the welfare dimensions to a few tractable ones. While most of the analyses have focused on monetary consumption (and/or income) vulnerability, this paper examines --in addition-- two other dimensions related to human capital investments: health and education. The second issue is how to operationalize the "propensity to suffer significant losses in welfare." A strand of the emerging literature has defined as ex-ante poverty risk. According to this view, a vulnerable household is defined as that for which the probability of having consumption below the poverty line is greater than some probability threshold. This threshold could be any value, but typically that of one half or the estimated poverty incidence is used (Chaudhuri, 2000; Chaudhuri, Jalan and Suryahadi, 2001; Dercon, Pritchett, Suryahadi and Sumarto, 2000). In most of these studies predicted outcome distributions rely on the assumptions about the random errors. Shocks are subsumed in the errors and all the information on risk (predicted probabilities) comes 6 from their processes. If one only counts with cross-section data, under some heavy assumptions, intertemporal shocks (such as falling into poverty) are predicted from cross- sectional variability. The estimation of such propensity, then, imposes many assumptions on the underlying distribution of welfare and is affected by household and time specific unobserved shocks. Dercon (2001), for example, criticized the cross-sectional approach arguing that the consumption distribution observed in a period of time is contaminated by unobserved components that could be idiosyncratic to the household (individual) or the period when the survey was carried out. The availability of panel data in developing countries has enabled researchers to examine the dynamic of observable dimensions of welfare, particularly consumption or income. Since households are observed over time, household specific components are controlled for. In addition, surveys implemented during crisis periods have allowed the identification of household behavioral responses to specific shocks (such as currency crises in Russia and Indonesia, or hurricanes in Central America). Distributional assumptions are less restrictive and the welfare distribution can be estimated from actual information about welfare and shocks experienced by households.3 An example of this approach is the series of papers exploiting the Indonesian panel data that assess the effects of the devaluation of the Indonesian currency on household outcomes. This paper follows this latter approach to examine the consumption/income vulnerability and other dimensions. 4. Evdennce Ifrom MoIdova The data in this paper is from the Moldova Household Budget Survey. This household survey has been conducted monthly by the Moldova Department of Statistics since 1997. The sampling frame is based on the pooling lists for the December 1996 presidential elections and is intended to be representative at the national level by quarter. It contains information similar to other household budget surveys (like the LSMS, e.g.), with some basic information at the person level, and much more detailed information at the household level. This paper exploits both the cross section and the panel components of the survey. Out of a quarterly sample of about 1,600 households, a sub-sample is followed with different patterns.4 For our analysis of consumption and expenditures in health, we form a panel using all households with two yearly observations, one before and one after the Russian Crisis (i.e., the third quarter of 1998). More specifically, we exclude the observations at 1998 Q3, and take as the pre-crisis data that for those households interviewed in 1997Q4, 1998Q1 and 1998Q2. Then the post-crisis data correspond to the second yearly 3 Also, some recent studies are starting to place nsk "structurally" in their model by conditioning on available nsk information in their prediction model A more detailed review of this sort of approach is in Dercon (2001). 4 There are three rotation schemes, one quarterly and two yearly. In the quarterly rotation, some panel households are surveyed twice in 2 consecutive quarters (i.e., they are interview for a second time after 3 months). In the yearly rotations, some others are surveyed once a year (during the same month) for 2 years, while some others are surveyed once a year (durmg the same month) for 4 years. 7 observation for these households during 1998Q4, 1999Ql and 1999Q2, respectively. The data we analyze contains 1766 households in three different dimensions: consumption, education and health. 4.1 Consumption changes during the crisis In terms of consumption, there was an overall deterioration in living standard for the period after the Russian Crisis. We can see this from Figure 4. This figure presents kernel density estimations of (log) consumption per capita for the pre- and post-crisis period. A (log) poverty line is also shown.6 The estimated distributions of (log) consumption show a shift to the left for the post-crisis period, increasing overall poverty. Indeed, poverty incidence increased by about 10% (from 52% to 62%). A decomposition of poverty indexes before and after the crisis by selected household characteristics is provided in the appendix. Desinty at t=O * Desinty at t=1 .8501 00043 ___________________ _ 2.02217 7 '474I Log Consumpton Figure 4 Beyond this aggregate distributional shift, the crisis presumably hit households at different income levels differently. If so, their relative ranking might have changed accordingly. The transition matrix in table 1 below is a first look at this effect. If the deterioration in living standard suggested in figure 1 were felt evenly across the whole population, then households in any given quintile group would remain in their same quintile group in the second period. Table 4.1 reveals that, indeed, the majority of the households in a given consumption quintile had moved to a different quintile after the crisis. The more "volatile" quintiles are the intermediate quintiles, for which only 28% of the households remained in their 5 Some possible outliers (households with the change in log consumption at 2 standard deviations away from the mean) were excluded. 6 The line is the log of the poverty line as calculated by Signoret and Murrugarra (2001). Consumption is measured as per-capita consumption m real Moldovan Lei. Price indexes used to deflate nominal consumption are estimated in the same source 8 initial quintile, while the top and bottom quintiles are relatively more stable, in this respect. Table 4.1. Transition Matrix (percentages) Post-crisis Pre-cnsis Quintile 1 Quintile 2 Quintile 3 Quintile 4 Quintile 5 Quintile 1 47 26 17 8 2 Quintile 2 27 28 21 18 7 Quintile 3 17 20 28 25 10 Quintile 4 5 17 22 28 28 Quintile 5 4 9 12 22 53 There are a significant number of households experiencing consumption drops at all level of initial consumption. Overall, about 63 percent of households in Moldova experienced a consumption drop. However, table 4.1 (last column) shows that the proportion of household experiencing consumption drops increases steadily as we move to higher quintiles of initial consumption. On average, the proportional change in consumption for the whole population is close to (and not statistically different than) zero. However, the proportional changes in consumption vary dramatically by initial consumption quintiles. Poor households in the lower quintiles experienced the highest proportional increases in consumption. Meanwhile, non-poor households in the upper quintiles experienced the highest drops in consumption. The information in table 4.2 points to initial consumption level as a key vulnerability covariate. It also seems to suggest that the main burden of the crisis was burden mostly by households that were initially non-poor. 9 Table 4.2. Per capita expenditure and the proportion of households with consumption drops by quintile Pre-crisis Post-crisis Change Prop. change Hhs w/ drops All sample. Mean 117.970 98.301 -19.670 -0.008 0.627 Std. err. 2.238 1.849 1.930 0.015 Median 96.081 80.027 -12.531 -0.151 Obs. 1766 Quintile 1. Mean 39.684 55.244 15.559 0.487 0.304 Std. err. 0.597 1.533 1.509 0.046 Median 42.311 49.464 9.464 0.260 Obs. 354 Quintile 2. Mean 68.896 71.659 2.763 0.051 0.582 Std. err. 0.397 1.955 1.982 0.030 Median 69.026 65.722 -4.810 -0.071 Obs. 353 Quintile 3: Mean 96.208 83.746 -12.462 -0.128 0.687 Std. err. 0.458 2.257 2.239 0.023 Median 95.834 78.775 -18.376 -0.195 Obs. 353 Quntile 4 Mean 132.797 112.583 -20.214 -0.144 0.722 Std. err. 0.724 2.887 2.962 0.023 Median 131.612 103.284 -29.809 -0.210 Obs. 352 Quintile 5 Mean 251.376 167.808 -83.569 -0.304 0.839 Std. err. 6.890 6.615 7.514 0.024 Median 213.353 142.771 -87.145 -0.409 Obs. 354 Tables 4.3 to 4.5 look at the consumption level before and after the crisis and the change and the proportional change in consumption, by certain household characteristics. From table 4.3, one can see that large cities (Chisinau, Beltsy) had the highest level of consumption in Moldova both before and after the crisis. The post-crisis measure on average is even higher than the pre-crisis measure outside large cities. By the same token, they also experienced the highest drop in consumption level, on average. However, proportional to their initial consumption level, it is the non-rural small towns the ones more severely hit. It is interesting to note the large difference between the mean and the median in this table (as well as in the following ones), suggesting a very asymmetric impact on consumption even after stratifying by country region (or other characteristics). In table 4.4, we look at the same variables by household size. This table shows that the smaller the household the larger the consumption drop and the proportional consumption drop. The table also shows that smaller households have higher per-capita consumption than larger households. The information in table 4.5, where we control for household head's education level, is less clear. In general, it seems to be the case that households with more educated heads have higher levels of consumption. But there is a non-monotonic relationship between 10 education and the consumption change or the proportional change. Households with a head with secondary education did better than household with a head with primary education. But household with a head with higher education had higher drops than household with a head with secondary education. Table 4.3. Per-capita expenditure by country regions Pre-cnsis Post-crisis Change Prop. change Large cities. Mean 165.468 136.693 -28.775 -0.046 Std. err. 7.683 6.215 6.840 0.029 Median 122.634 107.806 -17.205 -0.141 Obs. 368 Other towns: Mean 106.966 83.532 -23.434 -0.053 Std. err. 4.026 3.373 4.100 0.038 Median 89.277 67.654 -18.521 -0.214 Obs. 317 Rural: Mean 105.165 89.500 -15.665 0.016 Std. err. 2.073 1.764 1.762 0.020 Median 89.531 77.240 -11.153 -0.145 Obs. 1081 Table 4.4. Per-capita expenditure by household size Pre-crisis Post-crisis Change Prop.change Hh size 1: Mean 150.156 118.562 -31.594 -0.072 Std. err. 6.441 4.518 4.894 0.030 Median 120.757 96.684 -21.747 -0.196 Obs. 325 Hh size 2: Mean 131.010 107.562 -23.448 -0.029 Std. err. 5.325 3.801 4.468 0.029 Median 107.878 90.557 -17.773 -0.170 Obs. 446 Hh size 3 Mean 118.985 95.613 -23.371 -0.030 Std. err. 4.268 3.388 3.927 0.035 Median 99.739 77.530 -12.474 -0.152 Obs. 328 Hh size 4: Mean 100.109 91.324 -8.785 0.041 Std. err. 3.560 4.693 4.113 0.034 Median 83.692 73.340 -8.839 -0.125 Obs. 390 Hh size 5+: Mean 79.701 69.931 -9.769 0.061 Std. err. 2.780 2.668 2.900 0.043 Median 69.778 63.657 -7.962 -0.109 Obs. 277 Table 4.5. Per-capita expenditure by household head's education Pre-crisis Post-cnsis Change Prop. change Illiterate. Mean 109.903 93.196 -16.707 -0.063 Std. err. 8.815 6.530 8.147 0.068 Median 95.691 93.968 -19.966 -0.194 Obs. 47 Primary education: Mean 106.613 87.963 -18.650 -0.026 Std. err. 3.784 3.288 3.217 0.036 Median 95.307 76.984 -16.382 -0.191 Obs. 331 Secondary education. Mean 110.027 93.051 -16.975 0.006 Std. err. 2.365 1.922 1.993 0.018 Median 91.041 75.949 -10.988 -0.135 Obs. 1173 Higher education: Mean 182.320 145.275 -37.045 -0.049 Std. err. 10.751 9.152 10.321 0.045 Median 137.529 112.430 -21.084 -0.173 Obs. 215 The results in these tables have to be interpreted with caution. For instance, we saw in table 4 that small households suffer the most in terms of proportional change. It might be the case, however, that small households are more frequently observed in small towns where the proportional change in consumption is larger. Or that, as the same table 4 points out, smaller households have higher consumption levels, and households with higher consumption levels suffered larger drops (table 2). In figure 5. 1, we show non-parametrically the relationship between the change in log consumption (approximate in the limit to the proportional change in consumption) and initial log consumption. A vertical line at the poverty line is superimposed. Consistent with the story in table 4.2, the curve it is not flat, but negatively sloped. In figure 5.2 to 5.4, we show this relationship after stratifying for certain characteristics. Figure 5.2 shows that, after controlling for initial log consumption, households in other towns suffer the highest drops in log consumption, compared to other country regions. From figure 5.3, it seems that household head's gender does not make a big difference in the change of log consumption after controlling for initial log consumption. However, as previously wamed, figure 5.4 suggests that, after controlling for initial log consumption, it is large households (with 4 or more members) those that suffer the larger drops in log consumption. 12 Figure 5. Kernel smoother: Change in (log) consumption on initial (log) consumption. Figure 5.1 .4- 3 4 5 8 i Log Consumption at P0- Figure 5.2 ktre CRISS Other Towns .2- 0 & -.2- -.4 3 4 5 i Log Consumption at t0 13 Figure 5.3 Male Head o female Head .4 - .2 -.. E~~~~~~~~~ Log Consump0n at tO -.2- -.24 -.4~~ ~~ - 3 4 5 B 7 Log Consumption at t=O Figure S thus gives us evidence of different distributional impacts, especially in te.ns of initial consumption, country region and household size. It also points to the importance of multivariate analysis to disentangle confounding factors. In the next section we make use of multivariate analysis to look more carefully at the determinant of consumption changes after conditioning for certain important characteristics. Multivarzate Analysis. Here we follow a more flexible and robust approach of looking at vulnerability. We use quantile regression analysis, which requires very weak distributional assumptions, to look at exposure to consumption shocks. These 14 semiparametric estimations are quite robust to misspecification of the errors as they permit for non-normal, heteroskedastic and asymmetric errors. Quantile regressions are also robust to outliers in the dependent variable and less sensitive to outliers in the regressors than mean regression procedures. Moreover, by providing a family of conditional quantile functions, quantile regression offers a much more complete view on the effect of covariates at different points in the consumption change distribution. This complements least squares analysis by allowing us to see if the estimates of the various effects at the conditional mean are indicative of the size and nature of these effects, say, in the lower tail. For example, do rural households have larger consumption drops compared to urban households on average? And does this urban-rural differential attenuate or increase in the lower tail, where consumption drops are significant? That is, it allows us to estimate the marginal effect of a covariate on y, at various points in the distribution, not just at the mean The specific model that we consider regresses the difference in the log of per capita consumption, before and after the Russian crisis, on a set of household characteristics including: initial consumption level, region dummy, household composition, household head's characteristics and household access to formal and informal safety nets. A complete list of all variables used in the analysis and their means are in Table 6. Results We estimate the model by OLS and by quantile regressions at five quantiles: 0.10, 0.25, 0.50, 0.75 and 0.90. The mean predicted proportional change in consumption of the models corresponding to the first three quantiles correspond to a negative change in log consumption (-0.768, -0.427, -0.131); the last two quantiles to a positive change (0.179, 0.462). The mean predicted proportional change in consumption corresponding to the OLS model is close to the median regression (-0.134). The estimated parameters are in table 7. The results suggest that the response of the change in log consumption to changes in certain household characteristics differs substantially at the different quantiles. In the discussion below, we concentrate mostly in the model for the lower quantiles and the conditional mean, given our interest in understanding vulnerability to consumption losses rather than consumption gains. Consistent with the descriptive statistics, households in other towns, and to a lesser extend those in the rural area, are more vulnerable to consumption drops than household in large cities. 7The hypothesis test that the estimated parameter vectors from the quantile regressions are equal is easily rejected. This suggests that we do not have a location model. 15 Table 4.6. Means of regression analysis variables Pre-crisis Post-crisis Variable Mean Std. Err. Mean Std. Err. Large cities 0.208381 0.009668 0.208381 0.009668 Other towns 0.179502 0.009135 0.179502 0.009135 Rural 0.612118 0.011598 0.612118 0.011598 Age 49.69309 0.370771 50.07701 0.367876 Head female 0.339185 0.011269 0.339751 0.011274 Smgle 0.0453 0.00495 0.042469 0.0048 Marred 0.671574 0.011179 0.674972 0.011149 Separate 0.01812 0.003175 0.020385 0.003364 Widowed 0.199321 0.009509 0.193092 0.009396 Divorced 0.065685 0.005897 0.069083 0.006036 Illiterate 0.026614 0.003831 0.019819 0.003318 Primary education 0.187429 0.009289 0.17667 0.009078 Secondary education 0.664213 0.011241 0.681201 0.011092 Higher education 0.121744 0.007783 0.12231 0.007799 Farmers 0.060023 0.005654 0.074179 0.006238 Hired m agnculture 0.215742 0.009791 0.220272 0.009865 Hired in non-agriculture 0.330691 0.011198 0.326727 0.011164 Self employed 0.017554 0.003126 0.019253 0.003271 Pensioners 0.352775 0.011374 0.34145 0.011287 Other 0.023216 0.003585 0.01812 0.003175 Household size 2.999434 0.036502 2.964326 0.036289 Numberunder6 0.219139 0.012151 0.197056 0.011433 Numberaged6-14 0.524915 0.01951 0.517554 0.019199 Number aged 15-17 0.14496 0.009007 0.152888 0.009422 Number aged 18-25 0.304643 0.014684 0.278029 0.013903 Number aged 26-64 1.475085 0.020315 1.469422 0.020562 Number over 64 0.330691 0.014249 0.349377 0.014589 Number earners 1.214043 0.023182 1.265006 0.024637 Plot size 0.039778 0.001331 0.047014 0.001983 Fraction agriculture 0.880871 0.003537 0.890881 0.003197 House area 63.53398 0.658636 63.21155 0.653748 Fraction living area 0.705316 0.002964 0.702402 0.002726 Housing ownership 0 857871 0.008312 0.875991 0.007845 Yearquarter 1 0.403737 0.011679 0.403737 0.011679 Year quarter 2 0.199887 0.009519 0.199887 0.009519 Year quarter 4 0.396376 0.011643 0.396376 0.011643 Pre-cnsis consumption (log) 4.550428 0.015634 Post-cnsis consumption (log) . . 4.403619 0.015229 Agncultural society, dummy 0.342582 0.011296 0.313703 0.011044 Social benefits, dummy 0.289921 0.0108 0.215742 0.009791 Pnvate transfers, dummy 0.463194 0.011869 0372027 0.011505 Households 1766 1766 16 Household head's characteristics like age and gender are not significant in explaining consumption vulnerability. Marital status and education do, however. Household with single households are more vulnerable than household with married heads. Regarding education, household with heads with higher education experienced less consumption drop than households where the head had no higher education. In terrms of socioeconomic group, it seems that there is not much difference whether the household head belongs to a particular socioeconomic group. The exception is within those households that experienced the most dramatic drops (0 = 0. 10), where those hired in agriculture did somewhat better relative to those hired in the non-agricultural sector. Longer households seem to be more vulnerable to consumption drops (OLS). Also, household with a large number of non-infant kids (aged 6 to 14) were more vulnerable to the most dramatic drops in consumption (0 = 0. 10). And large number of children aged 15 to 17 is associated with significant drops in consumption (0 = 0.25). Altogether, this suggests that, for any given household size, that is, after controlling for household size, a large fraction of kids is associated with a higher vulnerability to consumption losses. Meanwhile, variables intending to capture for holdings, like housing area, the fraction of the housing used for living and housing ownership are associated with a lower proportional drop in consumption. Initial consumption, as suggested in the descriptive statistics, is critically associated with the proportional drop in consumption experienced in Moldova. Higher initial consumption is associated with a deeper consumption loss. Interestingly, variables intending to capture the effect of formal and informal assistance mechanisms do not play a significant role in any of the models. Being part of an agricultural society do not seems to play a role in helping against consumption drops. Neither do the receipt of social benefits, like unemployment benefits, pensions and social security. Private transfers enter significantly only in the 0.25 quantile and median regressions, although with a small coefficient and an unexpected negative sign. In sum, household more vulnerable to consumption losses could be characterized as non- poor households from outside large cities, especially from small towns; with a large number of members, and more to the point, with a large number of kids; and with household heads who are single, have no higher education and are employed in non- agricultural activities. 17 Table 4.7. Estimated OLS and Quantile Regression Coefficients. Dependent variable: change in log PCE. OLS 8=010 0=0.25 0=050 0=075 9=090 Other towns -02509 ** -03768 ** -0.3461 ** -01888 ** -02566 ** -0.2197 Rural -0.1783 ** -0.2558 * -0.2363 ** -0.1631 ** -0.1780 * -0 1370 Age 0 0109 0.0203 0.0208 0.0086 0.0055 0 0064 Age sq -0 0001 -0.0002 -0.0002 * -0.0001 0 0000 0 0000 Head female 0 0221 0.1323 0.0935 -0.0067 0.0018 -0 0599 Single -0 2184 ** -0 3736 ** -0.3309 ** -0.2376 -0 1055 -0 0323 Separate -0 1686 -0 7410 ** -0.3306 -0.1299 0.0341 -0 0239 Widowed -0 0387 -0 0696 -0 0067 0 0501 0.0232 -0 0025 Divorced -0.0743 -0.1006 -0 0518 -0.0487 -0.0486 0 0562 Illiterate 0 0222 0.1218 0.0496 0.0728 0.0226 -0 0565 Primary education -0 0179 -0.0603 -0 0347 -0 0088 -0.0653 0 0159 Highereducation 01820 ** 02504 0.1681 * 0.2073 ** 01307 * 01314 Farmers -0 0716 0 0087 -0.0696 -0.0560 0.0044 -0 0840 Hired in agnculture 0 0352 0.1704 * 0.0269 -0.0150 0.0384 0 0406 Self employed 0.0434 00098 0.1072 00256 0.0834 -0 1049 Pensioners -0.0740 -0.0283 -0.0924 -0.0782 -0 0508 -0 0677 Other 0 0512 0.2425 * -0.0263 0.0763 0 0791 -0.1844 Household size (log) -0 1320 ** -0.1125 -0.0992 -0.1320 -0 1621 ** -0 0238 Number under 6 -0 0413 -0.0323 -0.0616 -0.0335 -0 0567 -0.0235 Number aged 6-14 -0 0572 ** -0.1344 ** -0 0599 -0.0292 -0.0045 -0 0783 Number aged 15-17 -0.0184 -00590 -0.1105 ** -0.0218 0.0349 0.0308 Number aged 18-25 00227 00071 -0.0199 -0.0012 0.0640 * -00331 Number over 64 -0 0288 0.0472 0.0148 -0 0146 -0 0170 -0 1168 Number earners -0 0271 -0 0352 -0.0274 -0 0199 -0 0149 -0 0078 Plotsize 1.0362 ** 0.9010 0.6552 1 1043 ** 04900 1.0263 Fraction agnculture -0.0404 0.1696 0.0872 0 0365 -0.0516 -0.3690 House area 0.0009 * 0 0011 0.0008 0 0010 * 0 0013 * 0 0011 Fraction living area 0 2993 ** 0 4856 ** 0.2614 0 2573 0 1996 0.4116 Housing ownership 0 1195 * 0.2550 0 2111 * 0 0676 0.0027 01386 Year quarter I -0 1498 ** -0.0990 -0.1262 ** -0.1150 ** -0 1778 * -0.2344 Year quarter 2 -01342 ** -0.0372 -0.0342 -0.1008 ** -02037 ** -0.2164 Consumption (log) -0 5276 ** -0 5412 ** -0.5018 ** -0 5261 ** -0.5276 ** -0 5339 Agricultural society -0 0390 -0.0472 -0 0224 0 0215 -0 0447 -0 1194 Social benefits 0 0210 -0 0428 -0 0280 0.0237 0.0522 0.0864 Pnvate transfers -0 0447 -0.0688 -0.0844 ** -0.0569 * -0.0586 0 0096 Constant 2 1441 ** 0.9525 ** 1.4491 ** 2 1291 ** 2.7037 ** 2 8968 Left-out variables are dummies for large cities, married, secondary education, hired in non-agriculture, number aged 26-64, year quarter 4. (*) Sigmnficant at the 10% level. (**) Significant at the 5% level. Std. Err. (not shown) from 50 bootstrap repetitions. 18 4.2. Education dimensions of welfare and vulnerability The deep economic and fiscal crisis in Moldova also reduced education expenditures since 1996, but particularly between 1998 and 1999. Real public education expenditures were slowly increasing between 1994 and 1997 at about 5 percent per year, but in 1998 and 1999, real expenditures in education decreased by more than 30 percent each year (Tibi, Berryman and Peleah, 2002). This represented an increase in arrears that reached a peak of 70 percent of total expenditures by the end of 1999. The impact of decreased expenditures and especially arrears in education must have been reflected in quality of education since those were concentrated in salaries, heating and electricity. Did the overall deterioration in the economy and the decline of education expenditures translate into a loss of education welfare? Since the widespread public coverage of basic education in Moldova that includes the provision of basic education materials does not require households to make significant contributions, the level of expenditures in education observed in the survey is very limited. The incidence of expenditures is negligible regardless of the quarter of the year8 and the paper examines the enrolment patterns between the academic years of 1997/98 and 1999/2000. This section, then, examines the school enrolment dimension. Since the analysis distinguishes enrolment by age groups (or their corresponding levels of education), Table 4.8 provides an overview of the education process in Moldova. In this paper the relevant education levels (grades) examined are Primary (1-4), Lower Secondary (5-9), and after graduation from Lower Secondary children can attend either Upper Secondary (10-12) or Vocational-Technical schools (that varies in duration). Higher education is not examined in this paper given the low incidence of higher education in the survey Childreis eknrolment in school. One dimension of welfare corresponds to the household decisions of whether children are sent to school or not. Households' decisions may very depending on the age and gender of the child, household composition and these decisions may be affected by economic downturns. The HBS contains very limited information on education choices of the household. The Family Roster, however, indicates the level of education and the attending institution if attending. Using this information an enrolment measure was estimated.9 The survey, however, does not distinguish between Upper and Vocational Secondary current enrolments, only distinguishes those for those individuals that have finished their education. 8 The school year starts m early September. Holiday, months are July and August. 9 Additional infomiation on school characteristics is only available for a few quarters throughout the survey. Thls information was not included in the analysis. 19 Table 4.8: Education levels in Moldova Education Level Eligibility Duration Delivery facilities (ages) (Grades) Preschool 1-6 - Preschool institutions Primarv 7-10 1-4 - Secondary - Gymnasiums Lower 11-15 5-9 - General secondary schools Lower 11-15 59 ~~~~~~~(GSS) - Lyceums 16-17 10-11 - GSS Upper (old system) 16-18 10-12 - Lyceums (new system) Vocatio- Professional After nal Gymnasium Polyvalent After Between 3 and 6 Gymnasium years Colleges Gymnasium: Duration varies, Graduates from GSS (grade 1 lg depending on 11) and Lyceums (grade 12) Lyceums: 12g. grade admitted. Source: Tibi et al. (2002). Note: The Government of Moldova plans to upgrade the GSS to Lyceumns. Lyceums are located in Munucipalities and towns. Access to Lyceums in rural areas is very limited. Lyceum graduates get the Baccalaureate degree, required to enter University. The effects of economic crisis on children have shown differential effects depending on the age of children, gender and place of residence. An important analytical issue is whether to use panel or cross section data to examine enrolment since the panel sample would lead to confound the effect of the crisis with those of aging. The evidence from developing countries suggest that while the effects on younger children is negative because of delayed enrolment, among teenage children the effects are mixed: some point to increased enrolment due to low opportunity costs in the labor market, while other suggest a decreased enrolment to replace adult labor that goes to the market. Enrolment rates in Indonesia, a country whose enrolment rates were close to universal levels in the mid nineties, experienced some decline before and after the financial crisis in late 1997. Enrolment slightly decreased by 2 points among children 7-12 and it was more pronounced among girls, mainly associated with delayed entrance to school since drop out rates were relatively similar (Beegle et al., 1999). Enrolment for this age group also decreased more among children in the poorest quartile (between 5 and 6 points) and in rural areas (almost 4 points). For children aged 13-19 the evidence is less clear. Some survey data shows that enrolment decreased more especially among males from the poorest quartile and in urban areas, reflecting the need to participate in productive activities (Beegle et al., 1999), while another survey indicates that children of 14 or more were more likely to be enrolled probably because of diminished eamings opportunities (Frankemberg et el, 2001). Multivariate analysis indicates|that enrolment declines were 20 deeper for the poorest children and that in Indonesia, after achieving almost universal coverage, an income measure (per capita expenditures) became important again after the cnsis. The finding of relative improvements of enrolment among older children is corroborated by evidence from other countries. Schady (2002) uses cross sectional data from Peru to show that during crisis children do not decrease their enrolment because of lower opportunity costs (poor job market retributions). The Peruvian evidence shows, on the other hand, that the fraction of children studying and working decreased during crisis. Neri and Thomas (2001) show that household head movements from formal-to-informal labor have more deleterious effects than movements into unemployment in Brazil. Such movements raise the probability of a child entering the labor market, but only during growth periods. For children staying at school, such movements also increase the probability of repeating the school year, but only during recessions. What happened in Moldova? Since enrolment and most expenditure decisions are made at the beginning of the school year, this paper separates the sample in periods before and after the crisis. Periods before the crisis are quarters in the academic year 1997/98. The quarters corresponding to the year 1998/99 are difficult to characterize since the effects may have been observed once the school year had started. The year 1999/2000 is clearly a post-crisis years, despite the continued increased in poverty incidence. As other FSU countries, Moldova evidences an almost universal enrolment rate in its basic education level (grades 1 through 9). In the academic year 1999/2000 more than 95 percent of the children between 7 and 15 attended school (see Table 4.9). Enrolment is lower (about 55 percent) for children between 16 and 18, corresponding to Upper Secondary or Technical School. During the periods of analysis, net enrolment rates remained high or even increased for Basic Education. Preschool enrolment for children aged 6 years, however, declined partly due to the lower funding to this education level. Even enrolment for those directly facing labor market opportunities (aged between 19 and 20 years) increased during the worsening of economic conditions. The overall picture suggests that households were able to protect educational investments of their children (or that at the same time job opportunities were limited). Table 4.9:Moldova: Enrolment rates 1997-2000 (percent of children attending school) School level (ages) 97/98 98/99 99/00 Preschool (6) 11.5 8.5 7.5 Primary (7-10) 91.6 93.7 94.9 Lower Sec. (11-15) 96.5 96.0 96.9 Upper Sec. (16-18) 54.1 55.9 54.8 College (19-20) 18.7 25.1 22.2 Source: Moldova Household Budget Survey (1997-2000). 21 A rather different picture is observed when the analysis is detailed across urban and rural areas, particularly between large cities and other towns (Table 4. 10). Net enrolment rates improved during the school year 1998-99 for all education levels, except in rural areas for those aged between 11 and 18. Compared to increasing rates for urban areas, enrolment in rural areas slightly dropped, particularly for the 16-18 individuals (almost 3 points), although they recovered in the following year. Urban children are affected only in the 1999/2000 academic year, when enrolments for individuals aged 16-18 dropped almost 3 points, mainly in large cities. Table 4.10: Moldova: Enrolment rates 1997-2000 (percent of children of specific age groups attending school) Large cities 97/98 98/99 99/00 Preschool (6) 6.1 8.2 8.6 Primary (7-10) 93.2 94.2 95.4 Lower Sec. (I1-15) 97.1 98.2 99.0 Upper Sec. (16-18) 73.4 80.2 76.8 Other towns Preschool (6) 17.1 5.4 14.3 Primary (7-10) 85.5 89.4 92.8 Lower Sec. (11-15) 96.4 97.6 97.4 Upper Sec. (16-18) 62.1 68.8 67.5 Rural areas Preschool (6) 11.7 9.1 5.6 Primary (7-10) 92.6 94.7 95.4 Lower Sec. (11-15) 96.3 94.9 96.2 Upper Sec. (16-18) 44.9 42.0 44.6 Source: Moldova Household Budget Survey (1997-2000). Is there a gender dimension the changes in enrolment rates around the Russian crisis? Table 4.11 shows that, first, females have higher enrolment rates across different age groups in 19997/98. The differences are particularly higher for Preschool and Upper Secondary. These differences, however, were reduced during the crisis period mainly because of the overall increase in enrolments among males and lack of improvement among females aged I I to 18. Enrolment among girls was not significantly changed during the period but it was not increasing at the boys' pace either. Are these patterns the same for different socioeconomic groups? An examination of enrolment by consumption quintiles in Table 4.12 provides information about the groups that were affected by the worsening conditions after the Russian crisis. First, among children aged 7 to 10 years the overall increase in enrolment rates was not observed among the better off households while children from poorer households increased their chances by 2-4 percentage points (children from the top quintile actually decreased their 22 net enrolment between 97 and 98). As a consequence, the income gradient observed in 1997/98 is lost in 1999/2000. Table 4.11: Moldova: Gender and enrolment rates 1997-2000 (percent of children attending school) 97/98 98/99 99/00 Males Preschool (6) 9.4% 8.3% 5.0% Primary (7-10) 90.5% 92.5% 94.4% Lower Sec. (11-15) 95.6% 94.9% 96.8% Upper Sec. (16-18) 51.6% 54.9% 53.7% IFemal2es Preschool (6) 13.4% 8.8% 9.7% Primary (7-10) 92.7% 94.8% 95.5% Lower Sec. (11-15) 97.4% 97.2% 97.0% Upper Sec. (16-18) 56.6% 56.9% 56.0% Source: Moldovan Household Budget Survey (1997-2000). Table 4.12: Moldova: EnroDment rates by Quintiie 1997-2000 (percent of children attending school) Age group / Quintiles 97/98 9$/99 99/00 Age 7-10 91.6 93.7 94.9 Poorest 89.7 94.0 94.5 2 91.3 94.1 95.1 3 91.7 94.9 96.3 4 93.1 93.0 94.0 Wealthiest 93.3 91.3 94.5 Age 11-15 96.5 96.0 96.9 Poorest 94.9 92.2 93.3 2 97.1 96.3 96.4 3 95.8 97.5 98.0 4 98.4 96.4 99.0 Wealthiest 96.3 98.3 99.2 Age 16-18 54.1 55.9 54.8 Poorest 39.4 39.2 46.7 2 51.7 55.3 57.3 3 55.3 54.5 52.3 4 61.5 67.4 54.8 Wealthiest 64.5 71.7 65.4 23 A different story is observed among those aged 1 1-15: poorest children observed a reduction in their rates during the 1998/99 (crisis) year, but only to recover in 1999/2000. The rather flat income gradient in enrolment in 1997/98 is steeper in 1999/2000, when enrolment among the richest is more than 99 percent compared to 93 among children in the poorest quintile. Behind the almost constant enrolment rates for children with ages corresponding to Upper Secondary (16-18) there are very different socioeconomic patterns. While children from the poorest 60 percent evidenced an almost constant rate between 97/98 and 98/99, better off children evidenced a significant increase in their enrolment. In 1999/2000, however, the picture will change. The poorest children slightly increased their rates but the rates for the better off were significantly reduced, somewhat flattening the income gradient. The evidence poses some issues to be addressed in the multivariate analysis, table 4.13 to 4.15 shows the Probit results for the pooled sample and separate age groups. Since the better off quintiles were affected more by the crisis, it is consistent that enrolment did not increase for those aged 7 to 15. Why enrolment, then, increased significantly for those between 16 and 18, only to drop significantly in 1999/2000? Enrolment estimates for children 7 to 10 indicates that enrolment differentials across quintiles observed in 1997/98 were in fact associated with the consumption measure. Corroborating the evidence of decreased rates in rural areas, children from households headed by farmers have lower rates. In the next academic year (99/00) while plot size had a positive (wealth) effect on enrolment, the share dedicated to agriculture had a negative effect, still reflecting the worsening economic conditions in of rural areas. Regional differences became clearer (not shown). The regions of Beltsy and Cainari have had higher enrolment rates (compared to Chisinau region), but during the crisis these differences were more precisely observed. The steeper income gradient in enrolment among children aged 11 to 15 (corresponding to Lower Secondary) is corroborated by the increased importance of income and wealth (particularly housing). In this age group, household demographics play an important role, particularly after the crisis. Children (aged 11 - 15) in households with larger numbers of children below 6 are less likely to participate. Households could be protecting resources for children under 6 by not sending their elderly siblings to school (that is corroborated by the results for those aged 16 to 18), or that those 11-15 children were required to take care of their younger siblings to enable additional adult labor force into the market. The latter story has corroborative evidence if we assume girls were more likely to play such role, since female difference are observed in this age group. Despite the decrease in welfare among rural households, participation in Agricultural Societies helped the children in this age group to have higher enrolment rates in 98/99. Regional differences are also observed and the Beltsy and Cainari regions evidence higher enrolment, especially during the crisis year. 24 TablIe 4.13 Determinants of Enrolment for children aged between 7 and 10 years. (Estimates shown for the pooled sample and separate academic years) Pooled sample 97-98 98-99 99-00 dF/dx s.e. dF/dx s.e. dF/dx s.e. dF/dx s e. Age 0.0370 (0.0034) t 0.0409 (0.0073) t 0.0257 (0.0044) 00 0.0034 (0.0034) 4 Age squared 0.0000 (0.0000) 0.0000 (0.0000) 0 0000 (0.0000) 0.0000 (0.0000) Female 0.0010 (0.0045) 0.0081 (0.0075) 0 0060 (0.0073) 0.0003 (0.0007) Female head 0.0104 (0.0068) 0.0073 (0.0108) 0.0160 (0.0086) 0.0009 (0.0011) Single head -0.1624 (0.1122) ¢ -0.0941 (0.1090) Separated -0.0208 (0.0215) -0 0374 (0.0445) -0 0210 (0.0395) -0.0035 (0.0064) Widowed -0.0208 (0.0163) -0.0124 (0.0201) -0.0523 (0 0410) 0.0001 (0.0019) Divorced -0.0658 (0.0390) t -0.1885 (0.1171) 00 -0.0419 (0.0485) 0.0004 (0.0010) Illiterate -0.0530 (0.0584) -0.1126 (0.1915) Prnmary -0.0169 (0.0169) -0.0195 (0.0267) -0.0407 (0.0432) -0.0009 (0.0053) Higher + -0.0018 (0.0086) 0.0132 (0.0102) -0 0182 (0.0208) '° -0.0003 (0.0013) Farmers -0.0004 (0.0074) -0.0063 (0 0144) -0.0404 (0.0217) 0.0008 (0.0010) Hired Agnculture 0.0002 (0.0080) -0.0058 (0.0141) -0.0134 (0.0124) -0.0005 (0.0016) Self-employed 0.0017 (0.0145) -0.0368 (0.0571) 0.0005 (0.0008) Pensioners 0.0116 (0.0095) -0.0240 (0.0304) 0.0173 (00089) 0.0011 (00014) e Others 0.0078 (0.0150) -0.0110 (0.0348) Agncultural Society 0.0116 (0.0067) 0.0134 (0.0141) 0 0178 (0.0130) 0.0013 (0.0015) * Social benefits? -0.0146 (0.0089) -0.0114 (0.0152) 0.0006 (0.0140) -0.0047 (0.0065) Pnvate Transfers? 0.0004 (0.0047) 0.0050 (0.0080) 0.0088 (0.0076) -0.0013 (0.0017) log(PCE) 0.0064 (0.0041) 0.0129 (0.0070) 0 0.0010 (0.0067) 0.0002 (0.0006) log(household size) 0.0040 (0.0225) -0.0199 (0.0379) 0.0022 (0.0344) -0.0027 (0.0039) Number of age < 6 -0.0038 (0.0062) -0.0067 (0.0083) 00077 (0.0122) 0.0006 (0.0011) Number of age 6-14 0.0040 (0.0052) 0.0093 (0.0088) 00051 (00081) 0.0007 (00010) Numberofage 15-17 0.0037 (0.0070) 0.0021 (0.0110) 00119 (0.0115) 0.0013 (00021) Number of age 18-25 -0.0067 (0.0060) -0.0121 (0.0087) -0.0015 (0.0095) 0.0005 (0.0011) Number of senior 0.0076 (0.0071) 0.0153 (0.0100) -0 0029 (00090) 0.0005 (0.0012) Numberofeamer 0.0011 (0.0050) -0.0038 (0.0078) 0.0035 (0.0082) 0.0012 (0.0013) 0 Plot size 0.0000 (0.0000) 0.0000 (00000) 0.0000 (0.0000) 0.0000 (0.0000) * fraction in agnculture -0.0314 (0.0182) 0.0088 (00314) -00455 (00437) -0.0039 (0.0041) Total housing area 0.0000 (0.0001) 0.0001 (0.0002) 0.0002 (0.0002) t 0.0000 (0.0000) Share used for living 00219 (0.0194) 0.0130 (0.0279) 0.0575 (0.043i) -0.0023 (0.0035) Owner -0.0091 (0.0079) 0.0117 (0.0222) -0.0191 (0.0074) 00 0.0024 (0.0047) Winter 0.0105 (0.0053) 0.0126 (0.0100) 0.0229 (0.0069) -0.0002 (0.0008) Spring -0.0060 (0.0067) 0.0060 (0.0098) 0 0035 (0.0084) -0.0009 (0.0014) *0 Fall 0.0262 (0.0053) t 0.0182 (0.0093) 0 0311 (0.0071) 0.0018 (0.0019) 0 Other towns -0.0464 (0.0288) 4 -00211 (0.0243) -0 0116 (0.0328) -0.0121 (0.0134) Rural areas 0.0025 (0.0143) 0.0035 (0.0187) 00269 (0.0449) t 0.0013 (0.0032) Distance to school -0.0060 (0.0050) -0.0159 (0.0078) 0 -0.0007 (0.0079) -0.0001 (0.0006) Minutes to school -0.0001 (0.0003) 0.0004 (0.0006) -0.0006 (0.0007) ¢ 0.0000 (0.0000) 0* d9899* 0.0064 (0.0055) d9900* 0.0166 (0.0055) 00 Sample size 2982 1148 1004 830 Waldchi2(50) 319.09 312.26 809.6 495.12 Prob>chi2 0 0 0 0 Log likelihood -737.38 -230.694 -174.064 -106 587 Pseudo R2 0.2508 0.3206 0 2771 0 4358 Note: Other regressors included regional dummies (10) Standard errors shown are corrected for unknown heteroscedasticty and clustenng effects 25 Table 4.14 Deternminants of Enrolnent for children aged between 11 and 15 years (Estimates shown for the pooled sanple and separate academic years) Pooled sample 97-98 98-99 99-00 dF/dx s e. dF/dx s e. dF/dx s.e. dF/dx s.e Age -0.0093 (0.0012) ** -O 0076 (0.0017) ** -0.0093 (0.0025) ** -00075 (0.0022) ** Age squared 00000 (0.0000) 0.0000 (0.0000) 00000 (0.0000) 0 0000 (O 0000) Fen-le 00124 (0.0034) * 0.0086 (0.0039) * 0.0139 (00067) * 0.0019 (00043) Fernale head 0 0049 (0.0064) 0.0082 (00055) 0.0115 (O 0073) -0.0274 (0.0156) ** Single head -0.1044 (0.1014) -0.3503 (02408) Separated -0.0228 (O.0225) -0.0332 (0.0430) -0.1082 (0 0994) * -00012 (0.0127) Widowed -0 0308 (0.0168) * -0.0507 (0.0401) * -0.0407 (0.0294) 0.0040 (0.0072) Divorced -0.0386 (0.0239) * -0.1866 (0.1201) * -00257 (0.0317) 0.0035 (0.0078) Illiterate -0.0587 (0.0672) -0.1818 (0.2222) Prinniy 0.0012 (O 0104) -0.0123 (0.0221) 0.0051 (O 0161) 0.0076 (0.0046) Higher + 0.0153 (0.0049) * -0.0016 (0.0096) Fanners -0 0052 (O 0076) 0 0014 (0.0071) -0.0180 (O 0194) 0.0042 (0.0055) Hired Agriculture -00035 (0.0078) -0.0023 (0.0074) -0.0152 (0 0169) 0.0040 (0.0090) Self-eiloyed -0 0026 (0.0165) -0.0320 (0.0852) 0.0073 (0.0058) Pensioners 0.0081 (0.0079) 0.0139 (0.0045) -0.0028 (0.0190) -0.0002 (0.0148) Others -00813 (0.0528) * -0.1040 (0.0988) * -0.1663 (0.1197) * Agncultural Society 0.0064 (0.0063) 0.0047 (0.0066) 0.0179 (0.0115) * -0.0030 (0.0078) Socialbenefits9 -00087 (0.0068) -0.0136 (0.0112) -00022 (00136) -0.0130 (0.0116) Private Transfers? -0 0011 (0.0043) 0.0010 (0.0043) 0.0060 (0.0077) -0.0043 (0.0059) log(PCE) 00056 (0.0031) -0.0025 (00028) 0.0096 (00057) 0.0115 (0.0043)** log(household size) 0.0149 (0.0176) -0.0124 (00144) -0.0293 (0.0337) 0.0375 (0.0211) Nunber of age <6 -0.0122 (0.0049) ** 0.0083 (0.0059) -00035 (00079) -0.0184 (0.0069) * Nunber of age 6-14 0.0033 (0.0042) 0.0032 (0.0037) 00135 (00082) -0.0007 (0.0051) Nurmberofage 15-17 -0.0102 (00051) * -0.0040 (0.0043) -0.0029 (0.0084) -0.0045 (0.0052) Nunber of age 18-25 -00037 (O 0044) 0 0053 (0.0028) -00073 (O 0076) -0.0011 (0.0033) Numberofsenior -0.0041 (00060) 0.0021 (0.0057) -00048 (00122) -00108 (00064) Nurber of earner -0.0124 (0.0036) ** -0.0091 (0.0034) * -00077 (00065) -0.0159 (0.0052) ** Plot size 0 0000 (0 0000) 0.0000 (0.0000) 0.0000 (0 0000) 0.0000 (0.0000) fraction in agriculture 0.0136 (0.0129) -0.0451 (0.0293) 0.0462 (0.0427) 0.0169 (0.0123) Total housing area 0.0003 (0.0001) ** 0.0000 (0.0001) 0.0004 (O 0002) * 0.0004 (0.0001) ** Share used for livmg -00090 (0.0142) -0.0098 (0.0138) -00234 (0.0297) 0.0420 (0.0154) * Owner -0.0141 (0.0064) -0.0079 (0.0049) Winter 0.0027 (00065) -0.0187 (0.0108) * 0.0144 (00102) 00040 (0.0045) Sprng 0.0007 (0.0069) -0.0128 (0.0085) 0.0127 (0.0090) 0.0017 (0.0058) Fall 00083 (0.0058) -0.0169 (0.0121) 0.0187 (0.0086) * 0.0067 (00051) Other towns -0 0287 (00258) -0.0494 (00447) 0.0206 (0.0096) -0.9172 (01594) * Rural areas -0.0242 (0.0075) ** -0.0113 (0.0055) -0.0017 (0.0200) -0.0599 (0.0263) * Distance to school 0.0035 (00039) 0.0001 (0.0026) 0 0004 (0.0072) 0.0067 (0 0051) Minutes to school -0 0006 (0.0002) * -0.0003 (0.0002) -0.0007 (O 0005) -0.0004 (0.0003) d9899* -0.0023 (O 0058) d9900* 0 0065 (O 0050) Sanplesize 5475 1470 1321 1131 Wald chi2(50) 395.03 573.01 390 3 Prob > chi2 0 0 0 Log likelihood -769.21 -174.622 -200 876 -144.621 PseudoR2 0.1692 0.2454 02584 0.2719 Note. Other regressors tncluded regional dumrrmies (10) Standard errors shown are corrected for unknown heteroscedasticty and clustenng effects. 26 Table 4.15 Deterniinants of Enrolment for children aged between 16 and 18 years. (Estimates shown for the pooled sample and separate acaderrac years) Pooled sample 97-98 9S-99 99-00 dF/dx s.e dF/dx s.e. dF/dx s e dF/dx s e Age -0 0093 (0 0012) -0.0076 (0-0017) -0.3919 (0.0335) e -0.2129 (0.0383) e Age squared 0.0000 (0.0000) 0.0000 (0.0000) 0.0000 (0 0001) 0.0001 (0.0000) Female 0.0124 (0.0034) ee 0.0086 (0.0039) 0 0.0247 (0.0467) 0.0531 (0.0426) Female head 0.0049 (0 0064) 0 0082 (0.0055) -0 1662 (0 0681) 4 0.0778 (0 1050) Single head -0.1044 (0 1014) 0.3749 (0.1995) Separated -00228 (00225) -0.0332 (0.0430) 02486 (0 1804) -00881 (0 1302) Widowed -0.0308 (00168) -00507 (0.0401) o 02707 (0 1198) e -0.0722 (0 1317) Divorced -0.0386 (0 0239) e -0.1866 (0.1201) 4e 0.2137 (0 1493) -0.0859 (0.1579) Illiterate -0.0587 (0 0672) 0.2534 (0.2273) 0.0326 (0.2829) Pnrnary 0.0012 (0.0104) -0.0123 (0.0221) -0.3636 (0.1208) ¢ -0.1988 (0 1495) Higher+ 0.0153 (0.0049) e -0.0016 (0.0096) -0.0441 (0.1061) 0.0377 (0.0952) Farmers -0.0052 (0.0076) 00014 (0.0071) 0.0059 (0.1057) 0.0261 (00863) Hired Agriculture -0.0035 (0.0078) -0 0023 (0.0074) 0.0058 (0.0857) 0.0926 (0.0973) Self-employed -0.0026 (00165) -00320 (0.0852) -0.4195 (0.0919) e 0.2408 (0.1393) Pensioners 0.0081 (0 0079) 0.0139 (0.0045) -0.3469 (0 0929) e -0.4876 (0.0600) 4 Others -0.0813 (00528) -0 1040 (0.0988) ° -0.3963 (00870) e -0.4818 0.026338 00 Agricultural Society 00064 (00063) 0 0047 (0.0066) -0.2064 (0.0911) 0 -0.0244 (0.0797) Social benefits? -0 0087 (0.0068) -0.0136 (0.0112) -0.1573 (0 0805) 0.0468 (0.0715) Private Transfers? -00011 (0.0043) 0.0010 (0.0043) -0.1277 (0.0590) 0 -0.0830 (0.0533) log(PCE) 0.0056 (0.0031) -0.0025 (0.0028) 0.2371 (0.0475) e 0.0569 (0.0460) log(household size) 0.0149 (0.0176) -0.0124 (0.0144) 1.0638 (0.3102) e 1.1780 (0.3251) Numberofage<6 -0.0122 (0.0049) ee 0.0083 (0.0059) -0.2691 (00921) -0.3139 (0.0927) e Number of age 6-14 0.0033 (0.0042) 0.0032 (0.0037) -0.2666 (0 0729) 0 -0.3005 (0.0945) eo Number of age 15-17 -0.0102 (00051) * -0.0040 (0.0043) -0.2336 (0.0718) 00 -0.1674 (0.1052) Number of age 18-25 -0.0037 (0.0044) 0.0053 (0.0028) -0.1461 (0.0772) -0.1242 (0.0911) Number of senior -0.0041 (0.0060) 0.0021 (0.0057) -0.1538 (0.1012) -0.1807 (0.0965) Number of earner -0.0124 (0.0036) e -0.0091 (0.0034) 0 -02037 (0.0348) 0e -0 3861 (00483) 00 Plot size 0.0000 (0.0000) 0 0000 (0.0000) 0.0000 (0.0000) 0 0.0000 (0.0000) fraction in agriculture 0.0136 (00129) -00451 (0.0293) 0.1592 (0.2706) -0.1044 (0.1390) Totalhousingarea 0.0003 (00001) ee 0.0000 (0.0001) 00020 (0.0011) -0.0001 (0.0009) Share used for living -0 0090 (0.0142) -0.0098 (0.0138) -0.3684 (0.2143) -0.2617 (0.1952) Owner -0.0141 (0 0064) -0.0079 (0.0049) -0.3078 (0.0678) e 0 2865 (0.1351) Winter 0 0027 (0.0065) -0.0187 (0.0108) 0 0.0253 (0 0721) -0.0366 (0.0694) Spnng 0 0007 (0.0069) -00128 (0.0085) 0.0978 (0.0803) 00112 (00825) Fall 0.0083 (0.0058) -0.0169 (0.0121) 0.0477 (0.0746). -0 0030 (0.0754) Other towns -0.0287 (0.0258) -0.0494 (0.0447) -0.0475 (0.1557) -0.1851 (0.1843) Rural areas -0.0242 (00075) 04 -0.0113 (0.0055) -0.2967 (0.1340) 0 -02037 (0.1939) Distance to school 0.0035 (0.0039) 0.0001 (0.0026) -0.0254 (0.0565) 0.1023 (0.0609) Minutes to school -0.0006 (0 0002) -0.0003 (0.0002) -0.0027 (0.0041) -00051 (00039) d9899* -0.0023 (0 0058) d9900* 0 0065 (0 0050) Sarmple size 2575 723 669 643 Wald chi2(50) 731 25 647.59 1250 7 448.24 Prob > chi2 0 0 0 0 Log likelihood -1283 45 -174.62 -283 91 -307.01 Pseudo R2 0.2809 0.2974 0 3876 0.3111 Note: Other regressors included regional dumnumes (10). Standard errors shown are corrected for unknown heteroscedasticty and clustenng effects. Evidence for individuals aged 16 to 18 shows that crisis effected a reduction in enrolment in 1998/99. Children from households with pensioners as heads were less likely to be enrolled during the crisis and even less after. Household demographics play a more important role in school enrolment decisions. The evidence of younger siblings having a negative effect on the enrolment of their elderly is found again, and that the number of 27 children below 6 and those between 7 and 14 are equally important in reducing enrolment. While the hypotheses of intra-siblings competing resources and household labor demand cannot be distinguished here, the evidence suggests that teenagers with younger siblings were affected. This evidence is corroborated by the estimates of age effects. While the probability of enrolment decreases with age (especially among Secondary aged children) during and after the crisis, the probability of enrolment decreased significantly more, particularly for those between 16 and 18 years.10 Conversely, the number of earners is negatively associated with enrolment of teenagers, supporting the story of teenagers freeing household labor supply in times of crisis. In summary, during the period that covered the effects of the Russian crisis in Moldova, income and wealth played an important role in determining children enrolment in school. Distance and access to schools were less important. The results suggest that the crisis did have an effect on enrolment of children in the latter stages of Secondary Education, particularly those from families with younger siblings and pensioner heads. For younger children (7-10 years), however, income played a less significant role since poorer children increased their school enrolment while the better off stayed relatively constant. The evidence suggests that some decisions regarding household labor resources must have been made in order to provide additional labor supply (and additional income). 4.3 Health dimensions of welfare and vulnerability Moldova inherited an extensive healthcare system.'" This extensive legacy, however, has presented a major burden to the government in face of a decade of difficult transition and of a major regional crisis. The response has been one of major restructurings and expenditure cuts. Many healthcare units have been consolidated and public fiscal expenditure on health has been steadily decreased over the years. According to official statistics, in 1999, after the Russian Crisis, public fiscal expenditure on health was 357.6 millions of lei-2.9% of the GDP. This same figure was 537.1 millions of lei in 1997, or 6% of the GDP. In this section we explore the issue of vulnerability from the dimension of health. We first look at this from the aspect of household health expenditures and follow an analysis similar to that previously carried out for total expenditures. The idea here is to see the effects that the crisis might have had in households' health care budget allocations. Then, we turn to the related issue of health care utilization. Our major concern there would be to explore if, given the expenditure description, households in Moldova have seen their health care services utilization reduced after the crisis. '° The age effect also increases for those aged 11 to 15 during the cnsis, but in no significant fashuon. " As of 1998, all villages with 3,000 people or more were provided with a polyclinic and all smaller populations had some combination of health centers, posts, orfeldscher pomts. In total, the system consisted of more than 305 hospitals, 1,011 health posts, and 189 health centers, placing Moldova's delivery network above Region's averages (World Bank, 2002). 28 Health expenditures12 On the aggregate, expenditures on health decreased after the crisis when measured in levels or as shares of total expenditures. Health expenditures in Moldova during the period after the Russian Crisis drop on average by 7.4 lei per capita (a proportional drop of 47%). Likewise, health expenditure shares drop by about 3 percentage points, from 10.4% to 7.5% (a drop of 27.9%). These aggregate drops can also be seen from figure 3, where we have plotted estimated density functions for household health expenditure levels and shares. As for the general case of consumption, both distributions show a displacement to the left after the crisis. The following tables show, however, that this effect is not felt evenly across the population. To the contrary, the impact of the crisis on health care expenditures, follow an asymmetric pattern similar to that seen for total per capita consumption. Figure 6.1 Densty at tPO - Density at t-1 .263472 - .001535 _1.i4)315 ' ' 6.92762 Health Expenditures (log) 12 Because free access to basic services is widely available, the majority of households in our samnple report zero health expenditures for the last month, giving median values equal to zero for most of the population. We do not report proportional changes or median values in the following tables for that reason. 29 Figure 6.2 Density at t= * Density at t1- .275228- .001558 _ _ _ _ _X -8.6121 1.50382 Health Expenditure Shares (log) In table 4.16 we consider the effect by quintiles of initial consumption. The bottom quintile experienced an increased in health care expenditure levels of about 4 lei per capita (an increase of 158%). This quintile has no significant changes in its health care expenditure shares, however. The second and third quintiles have changes that are not statistically different than zero in either levels or shares. Meanwhile, regardless of whether we looked at it in terms of levels or budget shares, the two top quintiles experienced significant drops in health expenditures. Household in the top quintile expended on average 35.5 lei less per capita in health care after the crisis (a drop of 74%). Similarly, health expenditure shares for this top quintile drop by about 11 percentage points, from 18.2% to 7.1% (a drop of 61%). Table 4.16. Health expenditures by initial total expenditure quintile Shares Level Pre-crisis Post-crisis Change Pre-cnsis Post-crisis Change All sample Mean 0.104 0.075 -0.029 15 641 8.244 -7.397 Std. err. 0.007 0.006 0.008 1.342 0.766 1.447 Obs. 1766 Quintile 1 Mean 0.073 0.093 0.020 2.620 6.773 4.152 Std. err. 0.013 0.017 0.018 0.465 1.604 1.566 Obs. 353 Quintile 2 Mean 0.078 0.058 -0.020 5.364 5.485 0.121 Std. err. 0.012 0.014 0.017 0.805 1.756 1.891 Obs. 353 Quintile 3 Mean 0.073 0.076 0.002 7.236 7.337 0.101 Std. err. 0.011 0.012 0.015 1 101 1.411 1.752 Obs. 353 Quintile4 Mean 0.112 0.076 -0.036 14.930 9.161 -5.769 Std. err. 0.014 0.013 0.018 1.893 1.603 2.407 Obs. 353 QuintleS Mean 0.182 0.071 -0.111 47.915 12.390 -35.524 Std. err. 0.021 0.011 0.021 5.949 2.102 5.875 Obs. 354 30 Health expenditure levels are significantly higher in large cities than in any other country region. On the other hand, rural areas have the lowest health expenditure levels. The highest drop in health expenditure levels are also for households in large cities. This drop in large cities is of about 10.4 lei per capita. However, in terns of budget shares, it is outside large cities were the higher drops in health expenditure are found. Small towns and rural regions have drops in health expenditures shares of about 4 and 3 percentage points (drops of 33% and 34%), respectively, while in large cities this drop is only of (a not statistically significant) 1 percentage point (a drop of 10%). Table 4.17. Health expenditures by country region Shares Level Pre-crisis Post-crisis Change Pre-cnsis Post-crisis Change Large cities Mean 0.134 0.120 -0.014 24.819 14.430 -10.389 Std. err. 0.013 0.015 0.018 3.584 1.932 3.840 Obs. 368 Other towns: Mean 0.123 0.082 -0.041 16.670 8.821 -7.849 Std. err. 0.016 0.017 0.023 2.870 2.325 3.620 Obs. 317 Rural Mean 0.089 0.059 -0.030 12.334 6.047 -6.288 Std. err. 0.009 0.007 0.010 1.605 0.822 1.668 Obs. 1081 Table 4.18. Health expenditures by household size. Shares Level Pre-crisis Post-crisis Change Pre-cnsis Post-crisis Change Hh size I Mean 0.031 0.026 -0.005 7.166 3.356 -3.809 Std. err. 0.006 0.006 0.007 1.640 0.680 1.566 Obs. 325 Hh size 2: Mean 0.119 0.082 -0.037 21.125 9.311 -11.814 Std. err. 0.013 0.010 0.015 3.099 1.522 3.209 Obs. 446 Hh size 3 Mean 0.154 0.081 -0.073 25.418 9.644 -15.774 Std. err. 0.022 0.012 0.021 4.826 1 700 4.680 Obs. 328 Hh size 4 Mean 0.105 0.093 -0.012 12.113 10.319 -1.793 Std. err 0.014 0.017 0.021 2.039 2.233 3.029 Obs. 390 Hh size 5+ Mean 0.109 0.092 -0.017 10.421 7.809 -2.612 Std. err. 0.017 0.019 0.024 1.826 1.783 2.455 Obs. 277 31 Apart from single-member households, relatively small households have the largest drops in terms of levels and shares. But these households are the same with the highest health expenditure levels and shares before the crisis. And these may be the households with the highest total consumptions, which we know have large health expenditure drops (table 4.16). Shortly, in our multivariable analysis, we control for this to disentangle any confounding effect. Table 4.19. Health expenditures by household head's education Shares Level Pre-cnsis Post-cnsis Change Pre-crnsis Post-cnsis Change Illiterate- Mean 0.019 0.015 -0.004 1.446 1.690 0.244 Std. err. 0.013 0 008 0.015 0.909 1.049 1.249 Obs. 47 Primary education: Mean 0.091 0.053 -0.038 11.675 4.638 -7.037 Std. err. 0.014 0.010 0.017 2.179 1.182 2.427 Obs. 331 Secondary education Mean 0.105 0.081 -0.024 14.256 8.803 -5.453 Std. err. 0.008 0.008 0.011 1.501 0.984 1.665 Obs. 1173 Higher education Mean 0 139 0.094 -0.045 33.162 12.477 -20.686 Std. err. 0.020 0.016 0.022 6.527 2.717 6.716 Obs. 215 From table 4.19, we can see that health expenditures increase monotonically with household head's education. This is true in terms of both per capita expenditure levels and budget shares. Households with a head with higher education have the highest expenditures in health in Moldova. At the same time, it is this type of households the ones that see their expenditure on health decreasing the most. Regression analysis We run an OLS regression where the dependent variable is the change in health expenditure shares. Regressors include the household variables used in the analysis of consumption vulnerability plus variables to measure the time to get to both clinics and hospitals and the approximate distance to these institutions. As in our analysis for consumption, the inclusion of the year quarter dummies intends to control for any trend in expenditure shares (see a discussion on this in the consumption section). The results are in table 4.20. Coefficients for other towns and rural are both statistically different than zero, implying that health care expenditure shares drop more dramatically outside large cities, relative to 32 the changes in large cities (the omitted region). Although the coefficient for rural suggest a larger relative drop for the rural region than for other towns, we cannot reject the null hypothesis that the two regional coefficient are equal. Characteristics of the head like age, marital status, education or socioeconomic group seem not to play a role in the change in health expenditure shares. Household size plays an important role. Larger households have on average a larger drop in health expenditure shares. Specifics about the household composition do not seem to matter, however. Our variables intending to control for holdings (plot size and house are) are positively associated with the change. This seems to support the logic that having access to assets permits to cope against shocks. All else equal, relatively richer households have larger drops in health expenditure shares. None of the variables intending to capture the effect of formal and infor-mal assistance mechanisms (being part of an agricultural society, receiver of social benefit or private transfers) came out statistically significant. Distance to a hospital is positively associated with the change in health expenditure shares. A major distance to a hospital is indicative of the access to that type of service. Households farther away from a hospital will tend to have a lower utilization of hospitals and thus to allocate less of their budget to this use. Lower health expenditure shares for these households may explain their lower drops. 33 Table 4.20. OLS regression. Dependent variable: change in health expenditure shares. Coef. Std. Err Other towns -0.0776 0.0432 Rural -0.1248 0.0421 Age -0.0042 0.0055 Age sq. 0.0000 0.0001 Head female -0.0001 0.0267 Single 0.0279 0.0681 Separate -0.0691 0.0610 Widowed -0.0104 0.0393 Divorced -0.0092 0.0523 Illiterate 0.0522 0.0388 Pnmary education 0.0309 0.0266 Higher education -0.0023 0.0318 Farmers 0.0479 0.0474 Hired in agnculture 0.0059 0.0320 Self employed 0.0162 0.0981 Pensioners -0.0474 0.0424 Other -0.1788 0.0925 Household size (log) -0.1070 0.0428 Number under 6 0.0056 0.0286 Number aged 6-14 0.0238 0.0187 Number aged 15-17 0.0446 0.0310 Number aged 18-25 0.0351 0.0230 Number over 64 -0.0027 0.0217 Number earners -0.0244 0.0207 Plot size 0.3490 0.1867 * Fraction agncult 0.0466 0.0750 House area 0.0006 0.0003 * Fraction living area 0.0553 0.0722 Housing ownership 0.0055 0.0472 Year quarter 1 -0.0324 0.0199 Year quarter 2 -0.0311 0.0206 Consumption (log) -0.0720 0.0170 Agncultural society 0.0167 0.0245 Social benefits -0.0396 0.0248 Private transfers -0.0008 0.0195 Distance to clinic 0.0101 0.0135 Distance to hospital 0.0030 0.0012 Time to clinic -0.0005 0.0009 Time to hospital -0.0005 0.0006 Constant 0.5276 0.1949 R-squared 0.05 Left-out variables are duummes for large cities, rnamed, secondary education, hired m non-agnculture, number aged 26-64, year quarter 4. Std. Robust standard errors reported. (*) Significant at the 10% level. (**) Sigmficant at the 5% level. 34 Health care utilization Given lower public and private expenditures on health, healthcare utilization is of great concern. Indeed, official statistics suggests that health care utilization drops significantly in 1999. For instance, outpatient visits per person per year for most of the 1990s was consistently slightly above 8, and then fell sharply to just 5.6 in 1999. Also, like in most other country in the Region, in the 1990s there was a fall in the number of hospital beds. This fall, however, was particularly marked in 1999. In 1999, this number reached 825 beds per 10,000 people, down from 1,100 in 1998 and 1,150 in 1997. Other statistics like physician and hospital admissions per capita and the average length of hospital stays follow a similar trend (WHO, 2001). In this section we look at the micro evidence on health care utilization available in the household budget survey and explore the correlated associated with a lower utilization rate. Although the survey is not rich in health care utilization, it collects information on the number of visits per month to health clinics and hospitals and the approximate distance (physical and in time) to this institutions, and we exploit this data. Our data shows that clinics are far more commonly visited than hospitals in Moldova. From the 1,776 household in our sample, 560 households had at least one visit to a health clinic in the periods before the crisis. The number of households with, at least, one visit to a hospital is only 83, and 55 of these have also visited a clinic. Table 4.21 reveals that before the crisis 33.5% of the households in Moldova visited at least once a health clinic or a hospital per month. And most households in visited clinics (31.9%) compared to hospitals (5.2%). For the period after the crisis, the percentage of households that had visited either a clinic or a hospital decrease almost 7 percentage point, from 33.5% to 26.8%. This represents a significant drop in utilization of 20%, relative to the utilization before the crisis. Most of this decrease comes from a decreased utilization of clinics. Clinic utilization drops 6 percentage points, while hospital utilization drops by (a non significant) 1 percentage point. In table 13 to 16 show hospital utilization rates that do not exceed 7 percentage points, and for which the pre and post crisis values do not differ statistically from each other in any of the cases. For this reason we concentrate on discussing the utilization to clinics, understanding that this is where most of the action occurs. Dividing the population by quintile of initial expenditure shows that for all the quintile the post-crisis utilization rate is lower. However, except for the top quintile, utilization after the crisis cannot be found statistically different than utilization before the crisis. Standard errors across quintile are of about the same magnitude. But differences in utilization for the first 4 quintile are not large enough to make them significant, given the size of the standard errors. The top quintile's utilization drops to 28.4% after the crisis from a 40% prior the crisis-a significant drop of 31.5%. 35 Table 4.21. Health utilization by initial total expenditure quintile Pre-crisis Post-cnsis Visited Climc Hospital Visited Clinic Hospital All sample: Mean 0.335 0.319 0.052 0.268* 0.256* 0.043 Std. err. 0.011 0.011 0.005 0.011 0.010 0.005 Obs. 1766 Quintile 1: Mean 0.253 0.232 0.051 0.197 0.194 0.024 Std. err. 0.023 0.023 0.012 0.021 0.021 0.008 Obs. 353 Quintile 2: Mean 0.348 0.330 0.044 0.280 0.269 0.035 Std. err. 0.025 0.025 0.011 0.024 0.024 0 010 Obs. 353 Quintile 3: Mean 0 372 0.355 0.043 0.315 0.296 0.050 Std. err. 0.026 0.026 0.011 0.025 0.024 0.012 Obs. 353 Quintile 4: Mean 0.292 0.275 0.047 0.266 0.244 0.056 Std. err. 0 024 0.024 0.011 0.024 0.023 0.012 Obs. 353 Quintile 5: Mean 0.411 0.400 0.074 0.284* 0.274* 0.050 Std. err. 0.026 0.026 0.014 0.024 0.024 0.012 Obs. 354 (*) Statistically different (lower) than the corresponding pre-crisis value. In table 4.22, we divide the sample by country region. This table shows that the largest drops in health care utilization are outside large cities. The utilization drop in other towns is of about 8.5 percentage points (19.7%) and in rural areas of 7.1 percentage points (25.5%), while utilization in large cities drops by only 2 percentage points (6%). Due to the size of the standard errors, pre and post-crisis differences can be established statistically only in rural areas. Table 4.23 shows that health care utilization increases with household size. This suggests that more people in the household increases the propensity to someone in the household having to visit a health clinic. Comparing pre and post-crisis values, for households of all sizes, there is a decrease in utilization. The largest utilization decrease is for households with 4 members (9 percentage point, 25%). However, because pre and post values are largely not statistically different, table 15 does not provide a clear picture about household size and health care utilization changes after the crisis. 36 Table 4.22. Health utilization by country region Pre-crisis Post-crisis Visited Clinic Hospital Visited Clinic Hospital Large cities: Mean 0.356 0.349 0.036 0.335 0.328 0.019 Std. err. 0.025 0.025 0.010 0.025 0.025 0.007 Obs. 368 Other towns: Mean 0.436 0.432 0.036 0.349 0.347 0.029 Std. err. 0.028 0.028 0.010 0.027 0.027 0.009 Obs. 317 Rural: Mean 0.302 0.278 0.061 0.225* 0.208* 0.055 Std. err. 0.014 0.014 0.007 0.013 0.012 0.007 Obs. 1081 (*) Statistically different (lower) than the corresponding pre-cnsis value. Table 4.23. Health care utilization by household size. Pre-crisis Post-crisis Visited Clinic Hospital Visited Clinic Hospital Hh size 1: Mean 0.195 0.187 0.014 0.157 0.141 0.032 Std. err. 0.022 0.022 0.007 0.020 0.019 0.010 Obs. 325 Hh size 2: Mean 0.348 0.323 0.071 0.260* 0.249 0.042 Std. err. 0.023 0.022 0.012 0.021 0.021 0.010 Obs. 446 Hh size 3: Mean 0.365 0.365 0.046 0.309 0.298 0.038 Std. err. 0.027 0.027 0.012 0.026 0.025 0.011 Obs. 328 Hh size 4: Mean 0.383 0.354 0.062 0.28 1* 0.264 0.044 Std. err. 0.025 0.024 0.012 0.023 0.022 0.010 Obs. 390 Hh s,ze S+: Mean 0.386 0.371 0.060 0.359 0.350 0.063 Std. err. 0.029 0.029 0.014 0.029 0.029 0.015 Obs. 277 (*) Statistically different (lower) than the corresponding pre-crisis value. 37 Table 4.24. Health utilization by household head's education Pre-crisis Post-cnsis Visited Clmic Hospital Visited Clinic Hospital Illiterate Mean 0.170 0.170 0.000 0.063 0.063 0.040 Std err. 0.055 0.055 0.000 0.036 0.036 0.029 Obs. 47 Primary education: Mean 0.280 0.254 0.058 0.178* 0.165* 0.032 Std. err. 0.025 0.024 0.013 0.021 0.020 0.010 Obs. 331 Secondary education: Mean 0.349 0.334 0 055 0.295* 0.281* 0.051 Std. err 0.014 0.014 0.007 0.013 0.013 0.006 Obs. 1173 Higher education. Mean 0.386 0.372 0.036 0.312 0.304 0.019 Std. err. 0.033 0.033 0.013 0.032 0.031 0.009 Obs. 215 (*) Statistically different (lower) than the corresponding pre-cnsis value. In table 4.24, we at health care utilization by the household head's education. The table shows that household utilization increases with the level of education of the head. For instance, 33.4% of households with a head with secondary education had visited a clinic before the crisis, compared to 25.4% for households with head with primary education. This relationship would remain after the crisis. It also seems to be the case that household with more educated head had lower health utilization drops after the crisis. Although utilization decrease on average for all household, regardless of the head's education, household with head with secondary education had an utilization drop of 5.3 percentage points (16%), while households with head with primary education had a drop of 8.9 percentage points (35%). This seems to hold true for the other educational categories, with households with illiterate heads experiencing the largest drops in utilization and households with heads with higher education having small decreases. However, the imprecise means, due to the limited sample sizes for these other groups, impede to establish this significantly. All this may reflect that at higher education level there is more awareness on health. But it may well reflect that household with more educated heads are relatively richer or are located in relatively more urbanized area, which facilitate access to health care. In the next section we take a more careful look at this controlling for all these effects. 38 Regression Analysis We look at the factors associated with visiting a health clinic. For this, we run OLS regressions for before and after the crisis where the dependent variable is the household's number of visits to health clinics per month. Regressors are the same as those used for the part of health expenditures. The results are in table 4.25. A probit analysis on health clinic utilization provides similar insights to those given here and is included in the appendix. Consumption enters significantly in botLi equations and with a positive sign. This says that households with higher consumption visit health clinics with a higher intensity. This role to consumption gets larger in the equation for the post-crisis period. This make sense in a setting were public expenditures in health were falling down so rapidly. After the crisis, the variable for housing area (intending to capture for assets) and that for social benefits start to enter significantly, also reflecting this major reliance on household income for health care utilization. Distance to clinic and distance to hospital reflect monetary or nuisance costs of accessing these institutions, apart from any fee associated with the usage. Before the crisis both variable enter significantly in the estimated regression. Distance to a clinic enters with a negative sign. That is, clinic utilization is used less intensively the farther away the household is from the clinic. Distance to a hospital, however, enters with a positive sign. Since some health services can be obtained from either a hospital or a clinic (especially those of primary care), it is possible to see some substitutability between these two institutions. A higher distance to a hospital thus increases the intensity of health clinic utilization, reflecting a "cross-price" effect. For the post-crisis, the distance-to-hospital effect remains. Being close to a hospital ,implies a lesser utilization of health clinics. But the distance-to-clinic effect loses importance after the crisis. Seemingly, in this period of distress, the income effect previously discussed is so strong as to make this price effect loose importance. 39 Table 4.25. OLS regression. Dependent variable: Number of clinic visits. Pre-crsis Post-crisis Coef. Std. Err Coef. Std. Err Other towns 0.1523 0.3114 -0.0012 0.2523 Rural -0.4927 0.2850 * -0.2966 0.2333 Age 0.0208 0.0206 0.0231 0.0145 Age sq. -0.0002 0 0002 -0.0002 0.0001 Head female 00342 0.1116 0.2294 0.1628 Single -0.4094 0.1961 ** -0.1208 0.2052 Separate 0.1276 0.6226 -0.3988 0.2241 Widowed -0.3082 0.1536 ** -0.2800 0.1812 Divorced -0.3207 0.2227 . -0.4891 0.1835 Illiterate -0.2646 0 2088 -0.1261 0.2127 Pnmary education -0.1081 0.1564 -0.1606 0.0844 Higher education 0.4455 0.2551 * 0.0751 0 1333 Farmers -0.0405 0.1903 0.0983 0.1516 Hired in agriculture -0.0806 0.1582 0.1701 0.1792 Self employed 0.5200 0.4316 0.3567 0.5641 Pensioners 0.0758 0.2060 -0.0079 0.1767 Other 0.0868 0.3269 0.1210 0.2837 Household size (log) 0.1638 0.1957 -0.0003 0.1692 Number under 6 0.1298 0.1082 0.1524 0.1211 Number aged 6-14 0 0015 0.0762 -0.0167 0.0819 Number aged 15-17 0 0827 0.1269 0.2452 0.1586 Number aged 18-25 0.1037 0.0909 0.2046 0.1385 Number over 64 0.1758 0.1264 0.1197 0.1028 Number eamers -0 1864 0.0794 ** -0.0472 0.0823 Plot size -0 0502 1.0932 -0.2716 0.4068 Fraction agncult -0.3016 0.6136 -0.5062 0.5638 House area 0.0002 0.0018 0.0046 0.0023 Fraction living area 0.0387 0.4173 -0.4313 0.4458 Housing ownership 0.2190 0.1930 -0.2073 0.1938 Year quarter 1 0 1047 0.1044 0.0866 0.1053 Year quarter 2 -0.0048 0.1226 -0.0724 0.1229 Consumption (log) 0 1531 0.0822 * 0.2062 0.0738 Agncultural society 0.1574 0.1257 -0.0789 0.1161 Social benefits 0.1047 0.1230 0.2177 0.0998 Pnvate transfers 0 1046 0.1002 0.0564 0.0835 Distance to clinic -0.2050 0.0862 ** 0.0745 0.0671 Distance to hospital 0.0263 0.0104 ** 0.0199 0.0112 Time to climc 0.0038 0.0075 -0.0048 0.0037 Time to hospital -0 0032 0.0030 0.0036 0.0031 Constant 0.0144 1.0019 -0.3477 0.8578 R-squared 0.0728 0.0635 Left-out vanables are dumnues for large cities, married, secondary education, hired in non-agriculture, number aged 26-64, year quarter 4. Robust standard errors reported. (*) Significant at the 10% level. (**) Significant at the 5% level. 40 Table 4.26. Probit regression. Dependent variable: visit health clinic dummy. Pre-crisis Post-crisis Coef. Std. Err Coef. Std. Err Other towns 0.4389 0.1875 't 0.2242 0.1897 Rural -0.1422 0.1999 -0.1951 0 1939 Age 0.0056 0.0198 0.0109 0.0210 Age sq. 0.0000 0.0002 -0.0001 0.0002 Headfemale 0.0228 0.1170 0.0996 0.1137 Single -0.2274 0.2765 0.2487 0.2599 Separate -0.3823 0.3684 -0.0094 0.3210 Widowed -0.2314 0.1616 -0.1481 0.1683 Divorced -0.3925 0.2202 * -0.4679 0.2402 " Illiterate -0.2262 0.2749 -0.5278 0.3397 Primary education -0.1073 0.1266 -0.1860 0.1379 Higher education -0.0538 0.1429 0.0927 0.1380 Farmers -0.1622 0.1726 0.0541 0.1621 Hired in agnculture -0.1484 0.1332 -0.1820 0.1419 Self employed 0.3338 0.2993 -0.1481 0.3164 Pensioners -0.1592 0.1754 -0.1053 0.1920 Other 0.1516 0.3201 0.6477 0.3833 ' Household size (log) 0.6061 0.1819 ** 0.3400 0.2026 * Numberunder6 0.1086 0.0991 0.2327 0.1048 t Number aged 6-14 -0.0582 0.0742 -0.0198 0.0777 Number aged 15-17 0.0262 0.1127 0.0802 0.1120 Number aged 18-25 0.1041 0.0827 0.1003 0.0884 Number over 64 0.0557 0.0931 0.0289 0.0985 Numberearners -0.3016 0.0709 t -0.1003 0.0793 Plot size -1.2788 0.9063 -1.1399 0.5781 t' Fraction agricult. 0.2395 0.3410 -0.0162 0.3623 House area 0.0028 0.0015 * 0.0029 0.0016 ' Fraction living area -0.6916 0.3032 -0.4933 0.3634 Housing ownership -0.0787 0.1825 -0.0779 0.1900 Year quarter 1 0.0046 0.0829 0.1097 0.0877 Year quarter 2 -0.1444 0.1040 -0.0355 0.1084 Consumption (log) 0.3052 0.0657 4 0.2701 0.0730 t Agricultural society 0.1496 0.1073 0.0683 0.1175 Social benefits 0.0962 0.0963 0.3619 0.1086 ** Private transfers 0.0943 0.0785 0.0850 0.0841 Distance to clinic -0.1038 0.0682 0.0184 0.0660 Distance to hospital 0.0211 0.0047 't 0.0212 0.0049 t Time to clinic -0.0001 0.0044 0.0043 0.0042 Time to hospital -0.0023 0.0021 -0.0022 0.0020 Constant -1.9303 0.7213 t -2.2914 0.7418 ** Chi-squared 154.64 148.12 Left-out variables are dunmmies for large cities, married, secondary education, hired in non-agriculture, number aged 26-64, year quarter 4. (*) Significant at the 10% level. (**) Significant at the 5% level. 41 5. Summary The short-lived Russian crisis may have had long-term implications in those economies with strong economic links such as the trade partners. Moldova, one of the poorest countries in Europe, showed a relatively large dependency on trade, particularly with former Soviet Union countries (Russia). This feature exposed the Moldovan economy to the effects of the crisis in Russia that devaluated the Russian Ruble against the US Dollar. The change in relative prices (exchange rates) and the partial loss of the Russian market affected the Moldovan economy. This paper examines the question: How did households in Moldova respond to the crisis? What types of households were the most affected? Were investments in education and health delayed because of the crisis? Because of its linkage with trade activities, the Russian crisis affected those areas with strong dependence on trade activities. Since the major export product from Moldova is food and beverages (especially wine and wine products), rural areas and small towns (where some processing occurs) were particularly hit by the crisis. Although poverty in Moldova has been concentrated in rural areas, the changes after the crisis affected also the non-poor in urban areas (even large cities), increasing the overall poverty incidence from 52 to 62 percent. The larger negative impact on small towns is corroborated once controlling for other household characteristics. Moreover, living in a small town accentuated the negative effect on consumption. Even though the reduction in welfare (as measured by consumption) was widespread, some households were hit worse. Even though higher educated households were exposed to larger drops in consumption (since they were directly exposed to the externally driven shock), education of the household head did ameliorate the negative impact. Results on other covariates, such as household size or marital status of the head, reflect the urban- orientation of the shock since smaller households (most likely urban and single heads) suffered more than larger ones (rural and married heads). Other head's characteristics such as age and gender did not showed any association with increased vulnerability. Social safety nets did not reduce the impact on consumption. Neither social benefits from the government nor private transfers played a significant role in affecting the impact of the crisis on consumption. Only for those How did the worsening in consumption affect other dimensions of well-being? Household expenditures on education were already negligible to observe any change, but some systematic changes were observed in school enrolment. The most important impacts were observed for those children in Secondary school, and the evidence suggests that both household resources were scarcer and that household labor reallocation decisions were made. While rural areas observed a decline in Secondary School enrolment during the academic year 1998/1999, small towns and large cities observed a decline later in the following year and only for those aged 16 and above. The analysis showed that those declines were associated with income suggesting that economic conditions of the households were playing a more important role during times of crisis, even in a country with widespread educational coverage. The effects of household 42 demographics indicated that teenagers with younger siblings were less likely to be enrolled in school suggesting both that elderly children sacrificed for their younger siblings or that additional teenage labor was required at home to free adult labor up to the labor market. Similar to the consumption results, social assistance and social insurance mechanisms are not associated with better performance during the crisis. In education, household with pensioner heads experienced lower enrolment in Upper Secondary, also due to household labor decisions that involved the use of teenage resources. In contrast to findings from other countries, Moldova shows that crisis can marginally affect children enrolment, especially when household characteristics increase competition for resources and increases home labor demand for teenagers, and public expenditures experience significant declines. The impact on health was examined in two dimensions: expenditures and utilization. Health expenditures were significantly reduced particularly in small towns. Similar to what is found in consumption, health expenditures decreased more for households with heads with higher education (which were exposed to the largest declines in income/consumption). Households' assets like land ameliorate these negative effects but neither public nor private transfers have any offsetting effect. Given the broad coverage of public health services in Moldova, household responses may be reflected in utilization as well. About 34 percent of households utilized health services in Moldova, a fraction reduced to 27 percent after the crisis. Most of the utilization reduction was due to the decline in primary health care (not so much in hospital utilization), and mainly among the richest households and in rural areas. The parallel decline in public expenditures in health between 1997 and 1999 may explain part of the decline, but income and wealth variables have increased effects in periods of crisis. In contrast to the results on consumption and education, social benefits play a positive role in utilization of health care, particularly after the crisis. The (negative) role of distance to health care, is negligible when the crisis occurs, suggesting that distance and access are of lesser importance compared to financial and economic constraints. The analyses discussed in this paper show that Moldovan households where differentially affected by the impact of the Russian crisis, and that -- because of the nature of the crisis (fall in exports and devaluation) -- urban areas and the better off resulted bearing much more than the poor. Further separate analyses for urban and rural areas or by gender may shed additional light on the mechanisms underlying the decrease in welfare in Moldova. 6. References Beegle, K., E. Frankenberg and D. Thomas (1999) Measuring Change in Indonesia. RAND Labor and Population Program Working Paper Series 99-07. DRU-2014- WB/NIH. Deaton, A. (1997) The Analysis of Household Surveys. A Microeconometric Approach to Development Policy. Baltimore: John Hopkins University Press. 43 Deaton, A. and A. Fantozzi (1999) Prices and poverty in India. Research Program in Development Studies Working Paper. Princeton University. Deaton, A. and S. Zaidi (1999) Guidelines for Constructing Consumption Aggregates for Welfare Analysis. Research Program in Development Studies Working Paper No. 192. Princeton University. October. Dercon, S. (2001) Assessing Vulnerability. Mimeo. Jesus College and CSAE, Department of Economics, Oxford University. August. Frankenberg, E., K. Beegle, D. Thomas and W. Suriastani (1999) Health, Education, and the Economic Crisis in Indonesia. RAND draft. March. Thomas, D., K. Beegle, E. Frankemberg, B. Sikoki, J. Strauss and G. Teruel (2001) Education in a Crisis. RAND mimeo. March. Hentschel, J. and P. Lanjouw (1996) Constructing an Indicator of Consumption for the Analysis of Poverty. Principles ad Illustrations with Reference to Ecuador. LSMS Working Paper No. 124. World Bank. Washington, D.C. Leamer, E. (1988) "Measures of Openness," in Baldwin, R. (ed.) Trade Policy Issues and Empirical Analysis. The Chicago University Press. Moldova Economics Trend (2001) Economic Trends Quarterly Issue. July-September 2001. Neri, M. and M. Thomas (2001) Household Responses to Labor-Market Shocks in Brazil, 1982-99. Mimeo. Schady, N. (2002) The (Positive) Effect of macroeocnomic Crises in the Schooling and Employment Decisions of Children in a Middle-Income Country. World Bank Policy Research Working Paper No. 2762. January. Signoret, J. and E. Murrugarra (2001) Poverty Dynamics in Moldova. Europe and Central Asia Unit - Human Development Department. World Bank draft. Tibi, C., S. Berryman and M. Peleah (2002) Moldova's Education Sector. A Financing Strategy to Leverage System- Wide Improvement. Draft. February. 44 Appendix: Quantile regression The quantile regression methodology developed by Koenker and Basset (1978), and applied in the context of wage equations by Buchinsky (1994), among others, is a technique for estimating the 0-th quantile of a random variable. The quantile regression model assumes that conditional on a vector of characteristics, x, the 0-th quantile of the dependent variable, y1, is linear Qo (Yi I xl) = x'Pe which gives rise to a linear quantile regression model y, X'Pe + UOi where Qe ('., I xl) = 0. The coefficient vector P. is estimated by minimizing over be the expression E 0 I y, - x,be I + E(1- 0) I yx - x,b,I i y,.x,be i y,